Abstract

Developing physics-based models requires a detailed understanding of the important aspects of a system as well as many modeling-related factors. The trade between fidelity and efficiency requires abstractions of physical behavior that impose limits on the predictive ability of a model. The computational intensity of these models often requires a multifidelity approach, increasing the desire to understand the risks associated with each model. While traditional statistical techniques can identify irregularities in some model responses, determining the valid operating space for a non-linear, time-dependent, or stochastic simulation can be difficult due to both complexity and the amount of generated data. Identifying when a model fails to provide adequate response prediction is necessary for troubleshooting, verification, and analysis. Computational models often have difficulty reaching a stable solution, and even when a model runs to completion, the behavior can still be erroneous. Additionally, the region of applicability might typically be presumed to be a simple bounded area around the developed baseline, but some models have non-intuitive failures. It is easy for a human to visually verify the validity of a non-linear simulation response such as a stress-strain curve, but it is difficult for a computer program to do the same. Furthermore, when deviating from a baseline to perform uncertainty analysis or design space exploration, manual inspection quickly becomes impractical as the amount of generated data increases. This work seeks to develop a method using machine learning techniques of classification and clustering to provide automated detection of regions where predictive capability is lost. This is done by allowing the algorithm to compare trendlines and return recommendations for filtering prior to the application of traditional statistical or uncertainty quantification techniques. Automating this process allows for model developers and users to avoid tedious data processing steps and more effectively understand the limitations of a given model.